Letta Code was built by the team behind MemGPT (now Letta), a research project from UC Berkeley that demonstrated how LLMs could be given persistent, hierarchical memory beyond their context window. Letta Code applies this memory architecture to coding agents: where most tools start each session fresh, Letta Code accumulates learned facts about the codebase, team conventions, and task history across sessions, becoming more effective over time.
Key capabilities
Persistent agentic memory — Letta Code maintains long-term semantic memory (codebase facts), working memory (current task context), and episodic memory (past session history). The agent builds understanding of the project over weeks of use rather than treating each session as isolated.
Terminal-Bench performance — At launch, Letta Code ranked first among model-agnostic open-source agents on Terminal-Bench, the coding agent benchmark measuring autonomous task completion on real engineering tasks.
Multi-provider model support — Works with Claude, GPT-4o, Gemini, and local models through a unified provider abstraction. Model choice persists as a per-project preference.
Self-improving context engineering — The agent identifies when its context is becoming stale or irrelevant and actively manages its own memory window, reducing hallucination on long-running projects.
Autonomy level
Level 3 — Supervised agent. The agent edits files and executes commands under user direction, with memory persistence between sessions reducing the ramp-up time on each interaction.
Strengths
- Memory persistence gives a genuine productivity advantage on long-running projects
- Strong Terminal-Bench benchmark performance at launch
- Multi-provider model support with no vendor lock-in
- Apache 2.0 licence from a credible academic-to-startup team
- Improves autonomously as the agent learns project conventions
Limitations
- Memory persistence requires a Letta server (self-hosted or cloud), adding setup complexity versus simpler CLI tools
- Newer tool with less community documentation than Aider or Claude Code
- Memory-based approach is most valuable on longer projects; less advantage for one-off tasks
- 2.7K GitHub stars indicates a smaller community than leading tools